Research On Tomato Flower Target Detection And Location Algorithm Based On Deep Learning

In recent years,as different types of automatic picking robots continue to be used in actual operations,they have been widely accepted and used by the public.Picking robots during the crop harvest period can greatly reduce the labor workload of the operators,and during the crop cultivation period,intelligent spraying robots can also be used to assist the personnel in their work.Nowadays,the planting area of tomatoes is continuously expanding.After the tomato flowers are effectively identified,artificial pollination can not only increase the tomato fruit setting rate,but also prevent the fruit distortion rate.Therefore,adopting a fast and effective method to identify and locate tomato flowers is of far-reaching significance for improving tomato yield and quality.For greenhouses with large tomato planting areas,the artificial pollination rate is low,so robots are needed to help pollination to reduce the workload of operators.This research proposes a deep learning-based technology to identify and locate tomato flowers,in order to provide a test basis for the normal operation of the intelligent pollination robot.The main research contents are as follows:(1)Image collection based on tomato flowers.In order to improve the recognition accuracy,a total of20,000 images were collected in the greenhouse and the images after data enhancement and expansion.In order to accurately retain the original image information and ensure the recognition and positioning of tomato flowers in the field environment,this article does not perform image processing,and directly uses the original image as the training set and test set,and divides the tomato flower data set into three types of tags: unflowered Period,flowering period,and wither period.(2)Research on Tomato Flower Recognition Based on Deep Neural Network.The target detection algorithm uses the currently more popular Faster Rcnn and YOLOv4 algorithms,and the algorithm is modified according to the characteristic types of tomato flowers.The modified Faster Rcnn activation function is used to replace the Leaky Re LU function,and the FPN structure is added to improve the small target recognition ability of the algorithm.Replace a different backbone and use the upgraded Goog Le Net,which has its own BN layer inception_v2 and deeper ResNet101.At the same time,it compares and analyzes with the current YOLOv4 algorithm that integrates most of the optimization techniques,and selects the best recognition algorithm suitable for the data set in this article.Experiments show that the m AP value of YOLOv4 algorithm is 9.29% higher than Faster Rcnn.(3)Research on tomato flower positioning based on binocular vision.Use Mat Lab software to obtain the internal and external parameters,distortion coefficient and rotation and translation matrix of the binocular camera.Use the BM(Block Match)algorithm to obtain the corresponding depth information map,and then combine it with the trained recognition model to calculate the specific distance between the camera and the tomato flower.The test results show that the effective distance of the binocular camera is about 30 cm,the absolute error is within 1-5cm,and the relative average error is 7.48%.